Toward an AI-Automated Radiation Therapy Workflow: A Dual-Model Approach to Target Volume Contouring and Dose Optimization In Intermediate-Risk Prostate Cancer
Abstract
Purpose
Prostate cancer male reproductive system malignancy and the most common male cancer treated by radiation. Given the high incidence rate and complexities of treatment planning, artificial intelligence (AI) solutions in medical physics are essential to reducing healthcare system burden and improving access to patient-specific care. This study developed and validated two AI models to address resource-intensive bottlenecks in the care pathway: target volume (TV) contouring and dose optimization (DO), both of which are highly manual and vulnerable to inter-observer and intra-observer variations.
Methods
Two AI models were developed: 1) a TV segmentation model utilizing U-Net, a convolutional neural network architecture, and 2) a RapidPlan DO model with knowledge-based planning architecture. Both models were trained on 21 intermediate prostate cancer cases for risk-specific optimization. The TV model was validated by comparing its contours to human-expert contoured clinical target volumes using the DICE similarity coefficient. Performance of the DO model was evaluated on external beam radiation therapy protocols and cross-evaluated against clinical and outsourced models' plans.
Results
The TV model achieved a DICE score of 0.842 in a validation case, exceeding the benchmark average of 0.81 observed with TheraPanacea across 90 prostate cancer patients, despite facing significant challenges from extreme class imbalance (average of 0.0562% foreground voxels) and wide variation in prostate volume. The RapidPlan DO model demonstrated performance comparable to the clinical plan, even in a case where the outsourced model failed. However, model limitations were observed in patients with anatomical features outside the training dataset thresholds, highlighting the impact of data scarcity (n=21 compared to a typical n>60).
Conclusion
AI-driven automation for the medical physics workflow in the treatment of prostate cancer demonstrates high potential for decreasing time spent on manual, iterative tasks and reducing risks of plan quality variabilities while meeting the growing demand for efficient, patient-specific care.